[Triton] Shaoclee/085 triton fusion#690
Conversation
|
👋 Hi! Thank you for contributing to the vLLM project. 💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels. Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can either: Add 🚀 |
|
/lgtm |
this PR includes:
fused input_layernorm + per_token_fp8_quant
fused q_a_layernorm + kv_a_layernorm
tunned fused_qk_rope_cat_and_cache_mla (for bs=256, execution time goes from 24 ~ 27 us to 14 ~15 us)